A ConvNets-Based Method for Computational Intensity Prediction and Spatial Domain Decomposition

Objectives: The evaluation of heterogeneity is essential for domain decomposition in parallel geocomputation. According to the theory of spatial domain, the assessment of heterogeneity can be transformed to the computational intensity (CI) modelling, thus the key of domain decomposition is feature extraction and computational intensity prediction. However, the existing methods rely too much on expert knowledge for feature extraction and CI modelling of spatial domain, suffering from poor applicability of features, complex modelling process and low model accuracy. In order to relieve the dependence on expert knowledge and achieve accurate computational intensity prediction, a computational intensity prediction and decomposition method for spatial domain based on artificial intelligence (AI) deep learning is proposed. Methods: We use Convolutional Neural Networks (ConvNets) to capture the features of spatial domain automatically, and a fully connected layer is used to predict the computational intensity. A component was developed to match the spatial domain and the input of ConvNets. Results: Spatial intersection on vector data was implemented to compare the proposed method and traditional methods. The results demonstrated the advantages of the proposed method in terms of the usability and parallel performance. Conclusions: The proposed method optimizes computational intensity prediction and domain decomposition from data science perspective, which provides a reference on how AI deep learning can be used in high-performance geocomputation.